Data-driven models for building energy efficiency monitoring

Compartir
Nowadays, energy is absolutely necessary all over the world. Taking into account the advantages that it presents in transport and the needs of homes and industry, energy is transformed into electricity. Bearing in mind the expansion of electricity, initiatives like Horizon 2020, pursue the objective of a more sustainable future: reducing the emissions of carbon and electricity consumption and increasing the use of renewable energies. As an answer to the shortcomings of the traditional electrical network, such as large distances to the point of consumption, low levels of flexibility, low sustainability, low quality of energy, the difficulties of storing electricity, etc., Smart Grids (SG), a natural evolution of the classical network, has appeared. One of the main components that will allow the SG to improve the traditional grid is the Energy Management System (EMS). The EMS is necessary to carry out the management of the power network system, and one of the main needs of the EMS is a prediction system: that is, to know in advance the electricity consumption. Besides, the utilities will also require predictions to manage the generation, maintenance and their investments. Therefore, it is necessary to dispose of the systems of prediction of the electrical consumption that, based on the available data, forecast the consumption of the next hours, days or months, in the most accurate way possible. It is in this field where the present research is placed since, due to the proliferation of sensor networks and more powerful computers, more precise prediction systems have been developed. Having said that, a complete study has been realized in the first work, taking into account the need to know, in depth, the state of the art, in relation to the load forecasting topic. On the basis of acquired knowledge, the installation of sensor networks, the collection of consumption data and modelling, using Autoregressive (AR) models, were performed in the second work. Once this model was defined, in the third work, another step was made, collecting new data, such as building occupancy, meteorology and indoor ambience, testing several paradigmatic models, such as Multiple Linear Regression (MLR), Artificial Neural Network (ANN) and Support Vector Regression (SVR), and establishing which exogenous data improves the prediction accuracy of the models. Reaching this point, and having corroborated that the use of occupancy data improves the prediction, there was the necessity of generating techniques and methodologies, in order to have the occupancy data in advance. Therefore, several attributes of artificial occupancy were designed, in order to perform long-term hourly consumption predictions, in the fourth work. ​
​L'accés als continguts d'aquesta tesi queda condicionat a l'acceptació de les condicions d'ús establertes per la següent llicència Creative Commons: http://creativecommons.org/licenses/by/4.0/